Re: NetBSD hardclock overhead growth

[Johnny Billquist <bqt%update.uu.se@localhost>]

Hi.

On 2015-08-16 21:03, John Klos wrote:
> Hi,
>
> I was testing various versions of NetBSD on a VAXstation VLC to look at
> performance over time and attempting to characterize hardclock overhead.
> I have NetBSD 1.4, 1.5, 1.6, 6, 7 and current all NFS bootable. All were
> booted single user so no other daemons were running (just init and sh).
>
> It seems that code generation hasn't changed enough to be a factor in
> small synthetic benchmarks such as benchmarks/dhrystone.

Thanks! I have been wanting to do this for a long time... Another thing 
that would be interesting is checking file performance. bonnie or 
something similar should be good enough.

> Each compile was timed:
> time cc -O2  -DMSC_CLOCK    dhry_1.c dhry_2.c  -o dry2
>
> dry2 binary sizes:
> 1.4:    13801
> 1.5:    13822
> 1.6:    10687
> 6:    11968
>
> NetBSD 1.4 results:
>         69.05 real        54.71 user        10.61 sys
> Microseconds for one run through Dhrystone:  396.0
> Dhrystones per Second:                      2525.3
>
> NetBSD 1.5 results:
>         75.36 real        57.13 user        10.58 sys
> Microseconds for one run through Dhrystone:  427.0
> Dhrystones per Second:                      2341.9
>
> NetBSD 1.6 results:
>         97.40 real        80.31 user        10.98 sys
> Microseconds for one run through Dhrystone:  444.5
> Dhrystones per Second:                      2249.7
>
> NetBSD 6 results:
>        278.23 real       229.45 user        33.60 sys
> Microseconds for one run through Dhrystone:  472.2
> Dhrystones per Second:                      2117.7
> 1.6 binary:
> Microseconds for one run through Dhrystone:  465.6
> Dhrystones per Second:                      2147.8
>
>
> NetBSD 7 results (from NetBSD 6 binary since NetBSD 7 can't compile):
> Microseconds for one run through Dhrystone:  489.8
> Dhrystones per Second:                      2041.6
>
> NetBSD 7 with 1.6 binary:
> Microseconds for one run through Dhrystone:  489.7
> Dhrystones per Second:                      2042.1
>
> (NetBSD 7 panics when trying to run aout binaries)
>
> This isn't enough data to calculate hardclock overhead. Results from
> another machine would be needed. I will have my (repaired) VAXstation
> 4000/90a back soon and will test again with that machine.

Interesting. So we have just about tripled compile time, and the 
resulting files are slightly slower (or else system overhead have 
increased to make performance drop.

> Is anyone actively running NetBSD on an 11/780? :)

Not an 11/780, but anyway...
Krille:/usr/src# uname -a
NetBSD Krille.Update.UU.SE 5.99.44 NetBSD 5.99.44 (Krille) #387: Mon Jan 
24 18:15:12 CET 2011 
root%GW.SoftJAR.SE@localhost:/usr/obj/sys/arch/vax/compile/Krille vax
Krille:/usr/src# sysctl -a | grep hw
hw.machine = vax
hw.model = VAX 8650
hw.ncpu = 1
hw.byteorder = 1234
hw.physmem = 62914560
hw.usermem = 58363904
hw.pagesize = 4096
hw.disknames = ra0 ra1 ra2 ra3 ra4 ra5 ra6 ra7 ccd0
hw.machine_arch = vax
hw.alignbytes = 3
hw.cnmagic = \x27\x02
hw.physmem64 = 62914560
hw.usermem64 = 58363904
hw.iostatnames = ra0 ra1 ra2 ra3 ra4 ra5 ra6 ra7 ccd0 nfs0 nfs1
hw.ncpuonline = 1
Krille:/usr/src#

60 Meg, 2 Unibuses, 2*UDA-50 and 8*RA73, DELUA...
(Also one DMF-32 but I don't have that one configured in at the moment.)

	Johnny

--
Johnny Billquist                  || "I'm on a bus
                                  ||  on a psychedelic trip
email: bqt%softjar.se@localhost             ||  Reading murder books
pdp is alive!                     ||  tryin' to stay hip" - B. Idol